RepVGG-MEM: A Lightweight Model for Garbage Classification Achieving a Balance Between Accuracy and Speed

Abstract

Currently, existing garbage image classification models predominantly operate on low-end devices and encounter significant challenges, including limitations in computing resources, storage capacity, and classification accuracy. This paper proposes an improved lightweight model, RepVGG-MEM, specifically designed to address the resource constraints of low-end devices. The backbone of this model is derived from the lightweight RepVGG architecture, augmented by the integration of a multi-scale convolutional attention module to enhance high-quality feature extraction. Experimental results demonstrate that the RepVGG-MEM model outperforms its counterparts, achieving an accuracy of 93.26%, with a parameter count of 7.2 million and a floating-point operations (FLOPs) of 1.41 billion. This performance reflects a commendable balance between accuracy and processing speed. Furthermore, the model’s redundancy is minimized through pruning techniques, which significantly reduce both complexity and computational overhead. The optimal pruned version of the model is designated as RepVGG-MEM5. In this iteration, the parameter count is further reduced to 1.2 million and the FLOPs decrease to 0.55 billion, with a minor accuracy decline of only 1.17%. These findings indicate that it is possible to significantly reduce the model’s parameters and FLOPs without a substantial loss in accuracy, thereby enhancing the overall performance of the model while achieving an optimal balance of accuracy and speed. As a result, this research contributes to the development of an efficient and lightweight convolutional neural network model for garbage classification.